← All posts

Give Your AI Agent Its Own Linux Box (Without Giving It Your Laptop)

August 7, 2026 · 3 min read · The xShellz Team

TL;DR

  • An xshellz.box is a sandboxed, always-on Ubuntu 24.04 environment with fake-root via gVisor.
  • You connect over SSH and run any of six preinstalled coding agent CLIs from /root.
  • Persistent /root survives image resets, and the free tier has zero idle timeout.

You give an AI agent its own Linux box by provisioning an Agent Shell, an xshellz.box, via SSH. It's a hardened Ubuntu 24.04 sandbox where an agent gets fake-root powers through user-namespace remapping and gVisor, without ever touching your real machine. We run this same fleet infrastructure in production ourselves.

How can I give an AI agent a dedicated Linux environment to work in safely?

Provision an Agent Shell from xshellz. It's an always-on Ubuntu 24.04 LTS box with root transformed into a safe, isolated capability. Connect over SSH using the port and hostname from your dashboard, then run borg, claude, codex or any of the six preinstalled coding agent CLIs right from /root, which persists across image resets.

what is an agent shell and why do I need one?

An agent shell is a persistent, sandboxed Linux environment built specifically for AI coding agents. It's not a short-lived container that evaporates; it runs 24/7 even when your laptop is off. The box comes with six coding agent CLIs (borg, Claude Code, OpenAI Codex, Gemini CLI, opencode, pi), two Python agent frameworks (crewai, langgraph), and always-on agents like OpenClaw already installed. You log in as root and /root stays yours forever.

how do I set up an always-on box for an AI agent to SSH into?

Sign up at xshellz.com, choose a plan (free tier included, zero idle timeout), and your box spins up ready to SSH. The dashboard shows the exact connection string. Then just paste it into your terminal:

ssh -p <port> root@<host>
borg "fix the failing test"

Once inside, your agent can start with borg or any preinstalled CLI. The persistent home directory lets you store API keys, tmux sessions, and tool configs that outlive image updates.

is it safe to let an AI agent have root access to a Linux box?

Yes, because the root you give it is fake root. Every Agent Shell runs inside a gVisor sandbox with user-namespace remapping (userns-remap). The agent sees uid 0 inside the container, but outside it's mapped to a high, unprivileged uid on the host. The root filesystem is ephemeral; only /root survives. So even if an agent runs rm -rf / it can't touch the host or other tenants. See how gVisor isolates system calls and Linux user namespaces prevent real privilege escalation.

what's the difference between a throwaway sandbox and a permanent agent box?

A throwaway sandbox is for one-off jobs: you create it, run a task, and discard it within seconds. An Agent Shell is always-on and keeps your home directory and running processes alive between work sessions. Stopped boxes are held for 30 days before cleanup. That persistence makes the box suitable for agents that need a 24/7 presence, like an IRC bouncer, a long-running Hermes skill builder, or a Claude Code context file you update over weeks.

Frequently asked questions

Are coding agents and tools preinstalled, or do I bring my own?

Six coding agent CLIs are preinstalled: borg, Claude Code, OpenAI Codex, Gemini CLI, opencode and pi. They run right out of the box in Python 3 and Node 22. You bring your own API keys for each CLI. Nothing to compile or configure.

Can I host a website or blog from my agent box?

Yes. Choose any of 23 one-click apps (WordPress, Next.js, Hugo, Astro, and more). The app serves internally on port 8080, then you publish a frozen, malware-scanned copy live at https://<username>.turborg.com.

What happens to my data when the underlying image updates?

Only /root persists across a rebuild. Everything outside your home directory resets to a clean Ubuntu 24.04 LTS state. That's why we recommend storing agents, configs, and API keys under /root (which includes /home).